Project AURIONAI Use-Cases Research & Innovation
edge LLM

Gemma 4 E2B

Google's small Gemma 4 model for CPU and edge use. E2B-it has 337 derivatives, including Gemma-SEA-LION-v4.5-E2B.

GoogleGoogleUSAGoogle DeepMindGoogle DeepMindUK / USAGemma 4Family
Google
Levelmodel
Classedge LLM
Parameters5.1B
Context128,000
Released2 Apr 2026
Opennessopen weights permissive
Runs onlaptop, mobile or NPU, offline capable
Statusactive

Benchmark Results

maintainer-run
SEA-HELMSoutheast Asian and Indonesian
maintainer-run
64.13 95% CI [-1.39, +1.41]
04080

Score in score. Higher is better.

SEA-HELM Indonesian, open-weight models, average of 2,000 bootstraps18 Sep 2026 leaderboard.sea-lion.ai
maintainer-run
58.82 95% CI [-0.64, +0.66]
04080

Score in score. Higher is better.

SEA-HELM SEA average, open-weight models18 Sep 2026 leaderboard.sea-lion.ai
Where this comes from
Every score names who ran it: vendor-reported, maintainer-run or independent. A rank is shown with the population it was ranked in (gap rule 3).

Lineage

distillfinetune
Gemma 4 31B to Gemma-SEA-LION-v4.5-E2B-IT: distill, high confidenceGemma 4 E2B to Gemma-SEA-LION-v4.5-E2B-IT: finetune, high confidenceGemma to Gemma-SEA-LION-v4.5-E2B-IT: finetune, low confidenceGemmaGemmaGoogleGemma 4 E2BGemma 4 E2BGoogle, Apr 2026Gemma 4 31BGemma 4 31BGoogle, Apr 2026Gemma-SEA-LION-v4.5-E2B-ITGemma-SEA-LION-v4.5-E2B-ITAI Singapore, May 2026
AI SingaporeGemma-SEA-LION-v4.5-E2B-IT is a child (finetune)high confidencemodel card
Where this comes from
Parent and child relations from model cards, Hugging Face base_model fields and the genealogy tables of the research notes. The tree shows two generations each way, 4 models. Confidence and evidence are the notes' own.

R&D Topics

Low-resource and regional language resourcesused byQuantization and edge inferenceproduced byRegional or national foundation model training (continued pretraining at 7B to 70B)used bySmall language models (SLM)produced by

Runtimes

LiteRT (formerly TensorFlow Lite) and LiteRT-LMgoogle, ai-edge-litert 2.3.0, apache-2-0

Identity

Developer
Google, Google DeepMind
Family
Gemma 4
Released
2 Apr 2026 sourcesannouncement page
Weights published
–
Record checked
10 Oct 2026
Where this comes from
Developer and dates from the research notes. Release date is the vendor's public announcement; weights date is when open weights were published.

Architecture and Size

Architecture
dense transformer Dense with Per-Layer Embeddings; audio capped at 30 s
Total parameters
5.1B sources
Active parameters
2.3B
Checkpoint parameters
–
Note
2.3B effective (5.1B with embeddings)
Context length
128,000 tokens
Input
text, image, audio
Output
text
Where this comes from
Parameter counts as the developer states them. Where the checkpoint holds a different total, both are shown (register A16).

Openness and Access

Openness
open weights permissive
Licence
Apache License 2.0 sources
Access routes
open weights download
Where this comes from
The licence file governs over registry tags and aggregator labels.

Deployment

Runs on
laptop, mobile or NPU, offline capable
Footprint
Under 1.5 GB at 2-bit (Raspberry Pi 5); about 1.3 GB at Q4_K_M. Raspberry Pi 5: 133 tok/s prefill, 7.6 tok/s decode. Snapdragon 8 Gen 3: 20 to 35 tok/s (third-party estimate). Android via AICore Developer Preview. LiteRT-LM packages exist
Runtimes
LiteRT (formerly TensorFlow Lite) and LiteRT-LM
Where this comes from
Deployment tiers and runtimes as named in the notes. Footprint is the notes' figure, not a measurement by AURION.

Indonesia

Indonesian support
measured on a benchmark
  • SEA-HELM Indonesian leaderboard (ID aggregate, first score column of the Indonesian Competency and Task Scores table): 64.13 (95% CI -1.39 to +1.41), run by AI Singapore (SEA-HELM leaderboard maintainer) leaderboard.sea-lion.ai
Regional languages
not in the notes
Jakarta hosting
unknown
Languages
not in the notes
Where this comes from
Indonesia fields are carried on every model; unknown is an allowed value.

Industry Use Cases

No industry use case links this model.

Sources

  1. google/gemma-4-31B-it , Hugging Face huggingface.co https://huggingface.co/google/gemma-4-31B-it
  2. google (Google) huggingface.co https://huggingface.co/google
  3. aisingapore/Gemma-SEA-LION-v4.5-E2B-IT , Hugging Face huggingface.co https://huggingface.co/aisingapore/Gemma-SEA-LION-v4.5-E2B-IT
  4. SEA-HELM Leaderboard leaderboard.sea-lion.ai/detailed/SEA https://leaderboard.sea-lion.ai/detailed/SEA
  5. SEA-HELM Leaderboard leaderboard.sea-lion.ai/detailed/ID https://leaderboard.sea-lion.ai/detailed/ID
  6. जेम्मा रिलीज़  |  Gemma  |  Google AI for Developers ai.google.dev https://ai.google.dev/gemma/docs/releases
  7. Google Releases Gemma 4: Open Models Under Apache 2.0 datanorth.ai https://datanorth.ai/news/google-releases-gemma-4-open-models
  8. Gemma 4 E2B vs E4B: How to Run a Multimodal AI Model on Your Phone mindstudio.ai https://www.mindstudio.ai/blog/gemma-4-e2b-e4b-edge-models-phone-local
  9. huggingworld/gemma-4-E2B-it-litert-lm , Hugging Face huggingface.co https://huggingface.co/huggingworld/gemma-4-E2B-it-litert-lm
  10. Gemma 4: Byte for byte, the most capable open models blog.google https://blog.google/innovation-and-ai/technology/developers-tools/gemma-4/

To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Gemma 4 E2B", https://aurion.semai.id/models/google-gemma-4-e2b/, records of October 2026.